Make the memory GEP use the memory type

[Zentrik]

Given the comment LLVM is very bad at handling GEP with types different from the load seems like we want this. #54853 said this was changed in preparation for gep changing but it seems easy enough to make this change then.

Fixes #55090, closes #55107.

[Zentrik]

@nanosoldier runbenchmarks("array", vs=":master")

[vtjnash]

SGTM. I hoped llvm was getting better at this

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[vtjnash]

Looks like a very mixed bag of results? Some of those are known to be unstable, but it is unclear to me why this is quite so impactful looking

[Zentrik]

The simd regression disappears on llvm 17 but the sparse one does not. Perhaps this change should only be made for unions? If that doesn't fix the regressions, I guess there's not much to be done.

[gbaraldi]

It's so odd. The GEP i8 is the canonical way of doing this but what a mess.

[Zentrik]

Never mind, I must have got confused with some other issue, unions have nothing to do with this. Closing this then, we probably should merge #55107 even though it doesn't actually fix the regression.